Key facts
- Thinking Machines Lab has launched its first proprietary AI model, Inkling.
- Inkling is an open-weight model with 975 billion total parameters, drawing on approximately 41 billion for tasks.
- The model was trained on 45 trillion tokens of text, image, audio, and video, and reasons natively across these modalities.
- Inkling is designed to offer calibrated answers, flag uncertainty, and allow users to adjust 'thinking effort' for speed.
- The company's strategy emphasizes customizable AI for enterprises over one-size-fits-all proprietary models.
- Thinking Machines has a strategic partnership with Nvidia for computing infrastructure.
AI startup Thinking Machines Lab has launched its first proprietary AI model, Inkling, an open-weight system designed to be customizable by external developers and companies. This move challenges the dominant one-size-fits-all approach of major AI labs like OpenAI, Anthropic, and Google.
Inkling is a mixture-of-experts model with 975 billion total parameters, utilizing approximately 41 billion for specific tasks. It was trained on a diverse dataset of 45 trillion tokens encompassing text, image, audio, and video, and is capable of native reasoning across these modalities. The model aims to provide calibrated answers, flag uncertainty, and allow users to adjust computational effort for speed.
The company's core strategy is that AI models adapted by organizations will outperform those sold as finished products. This approach is gaining traction, with Microsoft CEO Satya Nadella warning that enterprises using proprietary models effectively pay twice due to subscription costs and the loss of proprietary business knowledge. Hugging Face CEO Clem Delangue also anticipates a shift towards private or open-source alternatives for most production AI work.
Thinking Machines highlighted a project with Bridgewater Associates, where a further-trained open-source model reportedly outperformed top proprietary AI models in financial reasoning tests at a significantly lower cost. The company emphasizes its rapid development cycle, reaching market and revenue in approximately nine months. While Inkling was partly trained using outputs from other open-weight models, future iterations will use fully self-contained post-training. The company has a strategic partnership with Nvidia for computing infrastructure and plans to generate revenue through its model-customization platform, Tinker.